What is an AI Agent, really?

2026-04-14 · 2 min read

Lately I keep hearing things like: automatically review PRs, automatically write tests, automatically generate release notes, automatically handle customer support…

It made me wonder — what does "Agent" actually mean?

Breaking it down

If you decompose a PR review agent, it looks roughly like this:

PR opened → webhook fires → AI runs review → posts comment

The "AI runs review" part is just calling an AI API with a set of instructions. Something like:

"Please review this PR, paying attention to naming conventions, security issues, and missing tests."

So the whole Agent breaks into two parts:

  1. Instructions: what to do and how to do it
  2. Trigger and integration: when to run and where to send the result

The instructions part

This is what I find most interesting.

Before the AI can do anything, you have to think through "how should this actually be done." For PR review:

These questions exist whether or not AI is involved. Good code review has always required answers to them. AI just forces the unspoken rules to be written down explicitly.

In a way, that's a good thing — it turns vague standards into words that can be shared and discussed.

The trigger and integration part

This is more traditional engineering work: wiring up webhooks, calling APIs, handling errors, routing results to the right place.

Not without value — but it's a separate concern from "how smart the AI is."

So what is an Agent?

My current understanding: an Agent is an AI call with clear instructions that fires automatically.

It's not something fundamentally new. It's more like automating something you used to do manually — except the thing being automated is "calling AI" instead of "running a tool."

Once you see it that way, a lot of things become clearer. There's no rocket science here — describe what you want clearly, write a good prompt, and that's the core of the whole thing. The trigger, the API wiring — those are just scaffolding around it.